BIOLOGICAL MODEL EVIDENCE · PART OF BAYES PHARMA.AI

Know what a biological AI model can—and cannot—support.

Find the exact version, inspect its evidence and limitations, compare context fit and preserve a reproducible evaluation record—without confusing availability with therapeutic validity.

Ask BioAtlasModels · evidence · benchmarks · verification · research

Find models Start an evaluation

82 scientific entitiesExact-version evidenceNo universal readiness score

EVIDENCE IS NOT A LEADERBOARD

BioAtlas asks not only “what did the model predict?” but “can we reproduce it, where can it fail, and should it change the decision?”

No fabricated universal score. Suitability depends on scientific task, evidence context, exact version, biological representation, calibration, failure boundaries, deployment constraints and context of use.

One system, three modes

Discover. Evaluate. Govern.

Start from a biological task or a known model. BioAtlas keeps context, exact versions and evidence boundaries visible through computational evaluation and handoff.

Discover

Find models, exact versions and evidence by biological task—not by leaderboard position.

Find models →

Evaluate

Compare context fit, reproduce claims, expose contradiction and identify where evidence stops.

Compare models →

Govern

Preserve the model-use question, validation plan, reviewed evidence and computational conclusion in one evaluation record.

Start an evaluation →

Connected model-evaluation path

Scope → Run → Review → Conclude.

Governed computation stays inside the evaluation evidence chain. Exact inputs, runtime identity, outputs and review state remain attached to every downstream claim and model-use recommendation.

Open evaluation workspace Inspect verification

01ScopeModel-use question, context and success criteria

02RunVersioned inputs, governed tools and durable computation

03ReviewArtifacts, claims, contradictions and qualification

04ConcludeModel-use recommendation and BayesPharma handoff

Scientific Verification OS

Executable Claim EngineTurn scientific statements into structured, rerunnable claims with population/context, intervention, comparator, outcome, evidence and reproducibility state.

Scientific Unit TestsRun golden cases, invariants, calibration checks, failure cases and out-of-domain tests against exact scientific model versions.

Bayesian Scientific Decision TwinRepresent programme beliefs explicitly and update them as evidence changes.

Governed execution

ToolsInspect scientific tool passports, intended use, exact versions, typed inputs and evidence boundaries.

JobsConfigure and monitor durable scientific execution with runtime state, cost context and reproducible artifacts.

BatchesRun governed candidate or dataset batches without losing project identity or per-item lineage.

Protein Structure Prediction AI Models AI Molecular Docking and Pose Prediction Models Protein Design AI Models Genomic Foundation Models Virtual Cell and Perturbation AI Models

Scientific intelligence

Explore models and evidence.

Search by model, organisation, capability or scientific principle. Catalogue fit is never presented as scientific validation.

82Matching records

Open full registry

Structure PredictionRecord

AlphaFold 2 / 3

Google DeepMind

The model that solved the 50-year protein-folding problem.

Structure PredictionRecord

RoseTTAFold / All-Atom

Institute for Protein Design, UW

The three-track network that followed folding into all-atom space.

Structure PredictionRecord

OpenFold

OpenFold Consortium / Columbia

A fully open, trainable reproduction of AlphaFold 2.

Structure PredictionRecord

ESMFold

Meta AI (FAIR)

Structure from a single sequence — no alignment required.

Structure PredictionRecord

Chai-1 / Chai-2

Chai Discovery

An AlphaFold3-class complex predictor, made freely usable.

Structure PredictionRecord

Boltz-1 / Boltz-2

MIT (Barzilay & Jaakkola labs)

Open-source AF3-quality structure — plus binding affinity.

Structure PredictionRecord

Protenix

ByteDance

ByteDance's open reproduction of AlphaFold 3.

Protein & Binder DesignRecord

RFdiffusion

Institute for Protein Design, UW

Diffusion models that hallucinate brand-new proteins.

Protein & Binder DesignRecord

ProteinMPNN / LigandMPNN

Institute for Protein Design, UW

Given a shape, design the sequence that folds into it.

Question → evidence → validation → handoff

Begin with the model-use question you need to answer.

Create one governed evaluation record for context, versions, computation, contradiction, review and the next evidence gate. Therapeutic programme authority remains in BayesPharma Discovery OS.

Start evaluation Compare models